特征融合和多峰检测的无人机目标跟踪

Feature fusion and Multi-Peak detection in UAV object tracking

无人机的目标跟踪在计算机视觉领域中是一个备受关注的研究热点.无人机跟踪目标在遭遇遮挡、尺度变化、光照变化等挑战会产生漂移,从而跟踪算法不能及时对模型进行更新.针对上述问题,提出特征融合和多峰检测的无人机目标跟踪算法,在特征融合的基础上采用多峰检测方法,结合置信度自动更新的响应策略方式,实现目标跟踪的精确定位.首先,提取目标的特征,并训练相应的滤波器模型,对特征进行加权融合,形成融合特征;其次,使用多峰检测方法对在目标中心附近图像块进行训练,采用简单有效的模型更新策略;最后,利用高置信度跟踪结果的反馈对模型进行更新,减少了模型漂移.在UAV123和VisDrone2019数据集上进行实验,本文所提的算法在跟踪目标遭受遮挡、光照以及尺度变化等影响下提高了跟踪的精确度和鲁棒性.

Object tracking of Unmanned Aerial Vehicle(UAV)is a piece of research concern in computer vision field.When the tracking object encounters the influence of occlusion,scale change and illumination variation,the tracking algorithm after drift cannot update the model timely.In order to improve the object tracking effect of UAV,a feature fusion and multi-peak detection algorithm in UAV object tracking is proposed.On the basis of feature fusion,the algorithm adopts multi-peak detection method and combines with the response strategy of confidence automatic updating to realize the accurate positioning of object tracking.Firstly,the features of the object are extracted,the corresponding filter model is trained.Multiple features are weighted and fused to form a fusion feature.Secondly,the multi-peak detection method is used to train the nearby image blocks around the object center.Finally,the model is updated with the feedback of high confidence tracking results.Experimental analysis and comparison are conducted on UAV123 and VisDrone2019 public datasets.The proposed algorithm improves the precision and robustness of tracking in the influence of occlusion,scale change and illumination variation of UAV.